Roles and Words in a massive NSSI-Related Interaction Network

Roles and Words in a Massive NSSI–
Related Interaction Network
Dmitry Zinoviev
Mathematics and Computer Science Department
Suffolk University, Boston MA
Presented at SunBelt 2019, Montreal CA

June 2019 INSNA SunBelt 2
What Is NSSI?
● Non-suicidal self-injury (NSSI), such as self-cutting or
self-burning, if the deliberate destruction of one’s
body tissue in the absence of suicidal intent.
● Approximately one in five of adolescents and one in
four of young adults in the USA have engaged in NSSI
(“self-cutters,” “self-burners”)

Where to Study Self-Harmers?
● Off-line: expensive, invasive
● On-line: cheap, noninvasive, in a naturally occurring
setting
– On LiveJournal:
● a blogging social networking site
● share skills and practices (especially concealment), ask
for help

Research Question
● Do NSSI-related topic starters and followers on
LiveJournal use different vocabulary, and if so, how do
they differ?

Research Strategy
● Build an interaction network of LiveJournal users
● Identify topics of discourse (ToDs)
● Find and explain the relationships between the
network attributes (such as centralities) and ToDs

Dataset
● ~140 NSSI-related thematic communities
● 15,678 active users
● 63,000 original posts
● 169,000 follow-up comments
● Posted in 2001–2012

Interaction Network Construction
● Interaction = response (comment) to the original post
or a comment
● A responds to B →edge from A to B
● Number of responses → weight of the edge
● Directed, weighted network
● 18 major network communities through Louvain
community detection
● Newman modularity 0.73

Network at a Glance
Node attributes
represent users’ roles

Interpretation of Attributes (I)
● In-Degree Centrality
– Author of requests for help or advice (topic starter), or
controversial statements
● Out-Degree Centrality
– Responder, advice-giver
● Closeness Centrality
– First responder (author of the first, or other lower-rank,
comment)
● Betweenness Centrality
– Mediator/broker

Interpretation of Attributes (II)
● Eigenvector Centrality
– “Important” member (in the most general meaning of the
term)
● Clustering Coefficient
– Participant of active multi-party discussions
We are not sure

Identify ToDs (I)
● Build a semantic network (a network of words). For
each post and comment:
– Remove frequent words (stop words)
– Lemmatize the remaining words
– Represent lemmas as network nodes
– Connect two words with an undirected edge if the lemmas
are at most five words apart in the text. The size of the
window is chosen to ensure that the resulting network is
neither too dense nor too sparse
– The number of co-occurrences is the edge weight

Semantic Network at a Glance

Identify ToDs (II)
● 11 major network communities through Louvain
community detection
● Newman modularity 0.37
● A community ↔ a collections of words that are
frequently used together ↔ a topic of discourse
● 11 major topics of discourse
● Each user has a vector of 11 topic memberships TDij
(∑i
TDij
=1)

Name ToDs
● Extracted semantic network communities (ToDs) do
not have names
● Name after the most frequent lemmas (e.g., “sad” →
the “sad” topic)
● Name via Amazon Mechanical Turk (* denotes “magic”
numbers)
– Select 25* most frequent words
– Submit to 25* AMT workers and ask to come up with a
single- or double-word name
– Accept the majority vote, if any

ToDs: Names and Top 7 Words
help Help, need, talk, stop, love, never, right
lifestyle Back, keep, away, around, put, stay, mind
friend Tell, friend, way, find, best, ask, mom
sad Sad, upset, depress, angry, depressed, pathetic
time Day, start, year, long, last, month, first
scar Scar, look, arm, leave, blood, enough, alone
hate Bad, life, hurt, hate, fuck, pain, feeling
rules Post, little, community, new, write, name, read
ana Yes, eat, disorder, depression, trust, etc., mental
s.i. Self, si, listen, sit, room, suicide, change
tools Use, razor, blade, word, cutter, usually, knife

Logit Regression
● Independent variables X:
– Six network centralities and the clustering coefficient (they
define the role of the user in the network)
● Dependent variables Y:
– Membership in each of the 11 topics of discourse (they
define the language use by the user)
– Binarized (above/below the median)

Significant Results
● Yellow nodes represent
independent variables
 Betweenness centrality is
not significantly related to
any Y
● Cyan nodes represent
dependent variables
● Arrows represent
statistically significant
(p≤0.05) relations between
independent and dependent
variables
● Thicker arrows represent
smaller p-values
● Blue arrows represent
positive coefficients
● Red arrows represent
negative coefficients

Interpretation
● Topic starters are not associated with “rules” and
“help”
● Responders are not associated with “tools,” visual
manifestations of NSSI (“scars”), and “time”
● First responders are not associated with “help,”
“friend,” and “time” – not a good sign!
● Intensive multi-party discussions related to “rules”
● Influence of propensity for brokerage is not significant
● The negative effect of eigenvector centrality on ”scars”
needs further research

Conclusion & Acknowledgment
● The structural roles and semantic preferences in an
NSSI interaction network are related.
● Topic starters and especially first responders
concentrate on negativity.
● Later responders concentrate on positivity.
● The author is grateful to the two anonymous reviewers
for their comments and inspiration

More NSSI Research from SU
● D Zinoviev, “Non-suicidal self-injury–related interests
in blogging social networks,” poster presented at
SunBelt, 2018
● D Zinoviev, D Stefanescu, G Fireman, L Swenson,
“Semantic networks of interests in online non-suicidal
self-injury communities,” Digital Health, 1, 2016

Roles and Words in a massive NSSI-Related Interaction Network

More Related Content

More from Dmitry Zinoviev (20)

Recently uploaded (20)

Roles and Words in a massive NSSI-Related Interaction Network